Cache filesystem fragmentation = squid death!

Hi,

We've had a couple of Squid deaths caused by fragmentation of the
cache_dir filesystems. A 'df' shows only typically about 80-85%
filesystem usage (a '-i' shows little inode usage - probably too low,
I'm increasing the bytes/inode).

'fsck' reports about 20% fragmentation (I presume that's internal frag-
mentation rather than external fragmentation, tho' since the # frag-
ments == the # free blocks!)

I wondered if anyone had any comments on this? What do other people
do for 'reserved space [-m]' and 'bytes/inode [-i]'?

I've 'newfs'd with '-m 10' and I set cache_dir to use about 95% of the
'used+avail' space rounded down a bit - the filesystems themselves
aren't filling, just fragmenting.

We're running 2.2S4+patches on Solaris 2.6 with normal UFS filesystems.
[Although I don't suspect that matters.]

Thanks,

  - Bob

-- 
 Bob Franklin <r.c.franklin@reading.ac.uk>     Tel. +44 (0)118 931 8432
 Systems and Communications, IT Services, The University of Reading, UK